[Feature] Move response jailbreak detection into a response-stage signal

config/config.yaml

@@ -422,6 +422,13 @@ routing:
         benign_patterns:
           - explain the policy
           - summarize the safety rules
+      # direction: response scores the model's output instead of the prompt, so
+      # this rule only exists once the model has answered; the selected
+      # decision's response_jailbreak plugin enforces on it.
+      - name: unsafe_completion
+        direction: response
+        threshold: 0.85
+        description: Detect jailbreak content in the model's own output.
 
     pii:
       - name: restricted_pii

config/fragments/signal/jailbreak/response.yaml

@@ -0,0 +1,12 @@
+# Response-direction jailbreak signal.
+#
+# Scores the model's output rather than the request, so it is only available
+# after the model has answered. The response_jailbreak plugin of the decision
+# selected for the request enforces on it; decision rules cannot read it.
+routing:
+  signals:
+    jailbreak:
+      - name: unsafe_completion
+        direction: response
+        threshold: 0.85
+        description: Detect jailbreak content in the model's own output.

dashboard/frontend/src/lib/dslSchemas.ts

@@ -3,6 +3,52 @@ import { getCapabilityPluginFieldSchema } from './dslCapabilityPluginSchemas'
 import type { FieldSchema } from './dslSchemaTypes'
 export type { FieldSchema } from './dslSchemaTypes'
 
+const JAILBREAK_SIGNAL_FIELDS: FieldSchema[] = [
+  {
+    key: 'method',
+    label: 'Method',
+    type: 'select',
+    options: ['classifier', 'contrastive'],
+    description: 'Detection algorithm',
+  },
+  {
+    key: 'direction',
+    label: 'Direction',
+    type: 'select',
+    options: ['request', 'response'],
+    description: 'request (default) scores the prompt; response scores the model output',
+  },
+  {
+    key: 'threshold',
+    label: 'Threshold',
+    type: 'number',
+    required: true,
+    placeholder: '0.9',
+    description: 'Minimum score to trigger (0.0-1.0)',
+  },
+  {
+    key: 'include_history',
+    label: 'Include History',
+    type: 'boolean',
+    description: 'Include conversation history in detection',
+  },
+  { key: 'description', label: 'Description', type: 'string' },
+  {
+    key: 'jailbreak_patterns',
+    label: 'Jailbreak Patterns',
+    type: 'string[]',
+    placeholder: 'Add jailbreak example...',
+    description: 'Contrastive mode: example jailbreak prompts',
+  },
+  {
+    key: 'benign_patterns',
+    label: 'Benign Patterns',
+    type: 'string[]',
+    placeholder: 'Add benign example...',
+    description: 'Contrastive mode: example benign prompts',
+  },
+]
+
 export function getSignalFieldSchema(signalType: string): FieldSchema[] {
   const policyFields = getPolicySignalFieldSchema(signalType)
   if (policyFields) return policyFields
@@ -263,44 +309,7 @@ export function getSignalFieldSchema(signalType: string): FieldSchema[] {
         { key: 'description', label: 'Description', type: 'string' },
       ]
     case 'jailbreak':
-      return [
-        {
-          key: 'method',
-          label: 'Method',
-          type: 'select',
-          options: ['classifier', 'contrastive'],
-          description: 'Detection algorithm',
-        },
-        {
-          key: 'threshold',
-          label: 'Threshold',
-          type: 'number',
-          required: true,
-          placeholder: '0.9',
-          description: 'Minimum score to trigger (0.0-1.0)',
-        },
-        {
-          key: 'include_history',
-          label: 'Include History',
-          type: 'boolean',
-          description: 'Include conversation history in detection',
-        },
-        { key: 'description', label: 'Description', type: 'string' },
-        {
-          key: 'jailbreak_patterns',
-          label: 'Jailbreak Patterns',
-          type: 'string[]',
-          placeholder: 'Add jailbreak example...',
-          description: 'Contrastive mode: example jailbreak prompts',
-        },
-        {
-          key: 'benign_patterns',
-          label: 'Benign Patterns',
-          type: 'string[]',
-          placeholder: 'Add benign example...',
-          description: 'Contrastive mode: example benign prompts',
-        },
-      ]
+      return JAILBREAK_SIGNAL_FIELDS
     case 'pii':
       return [
         {

dashboard/frontend/src/pages/ConfigPageSignalsSection.tsx

@@ -58,6 +58,126 @@ interface ConfigPageSignalsSectionProps {
   removeSignalByName: (cfg: ConfigData, type: SignalType, targetName: string) => void
 }
 
+function jailbreakDetailFields(
+  rawData: UnifiedSignal['rawData'],
+): Array<{ label: string; value: React.ReactNode; fullWidth?: boolean }> {
+  const fields: Array<{ label: string; value: React.ReactNode; fullWidth?: boolean }> = [
+    { label: 'Method', value: rawData.method || 'classifier', fullWidth: true },
+    {
+      label: 'Threshold',
+      value: rawData.threshold?.toString() || 'N/A',
+      fullWidth: true,
+    },
+    {
+      label: 'Direction',
+      value: rawData.direction || 'request',
+      fullWidth: true,
+    },
+    {
+      label: 'Include History',
+      value: rawData.include_history ? 'Yes' : 'No',
+      fullWidth: true,
+    },
+  ]
+  if (rawData.method === 'contrastive') {
+    fields.push(
+      {
+        label: 'Jailbreak Patterns',
+        value: (
+          <SignalStringListEditor
+            value={rawData.jailbreak_patterns}
+            onChange={() => undefined}
+            addLabel=""
+            emptyLabel="No jailbreak patterns."
+            itemLabel="Jailbreak pattern"
+            readOnly
+          />
+        ),
+        fullWidth: true,
+      },
+      {
+        label: 'Benign Patterns',
+        value: (
+          <SignalStringListEditor
+            value={rawData.benign_patterns}
+            onChange={() => undefined}
+            addLabel=""
+            emptyLabel="No benign patterns."
+            itemLabel="Benign pattern"
+            readOnly
+          />
+        ),
+        fullWidth: true,
+      },
+    )
+  }
+  fields.push({
+    label: 'Description',
+    value: rawData.description || 'N/A',
+    fullWidth: true,
+  })
+  return fields
+}
+
+type JailbreakFormState = Pick<
+  AddSignalFormState,
+  | 'jailbreak_threshold'
+  | 'jailbreak_method'
+  | 'jailbreak_direction'
+  | 'include_history'
+  | 'jailbreak_patterns'
+  | 'benign_patterns'
+>
+
+function jailbreakFormDefaults(): JailbreakFormState {
+  return {
+    jailbreak_threshold: 0.65,
+    jailbreak_method: 'classifier',
+    jailbreak_direction: 'request',
+    include_history: false,
+    jailbreak_patterns: [],
+    benign_patterns: [],
+  }
+}
+
+function jailbreakFormStateFrom(rawData: UnifiedSignal['rawData']): JailbreakFormState {
+  return {
+    jailbreak_threshold: rawData.threshold ?? 0.65,
+    jailbreak_method: rawData.method || 'classifier',
+    jailbreak_direction: rawData.direction || 'request',
+    include_history: !!rawData.include_history,
+    jailbreak_patterns: [...(rawData.jailbreak_patterns || [])],
+    benign_patterns: [...(rawData.benign_patterns || [])],
+  }
+}
+
+function buildJailbreakEntry(name: string, formData: AddSignalFormState): JailbreakSignal {
+  const jailbreak_threshold = formData.jailbreak_threshold ?? 0.65
+  if (jailbreak_threshold < 0 || jailbreak_threshold > 1) {
+    throw new Error('Jailbreak threshold must be between 0.0 and 1.0.')
+  }
+  const method = formData.jailbreak_method || 'classifier'
+  const jailbreakEntry: JailbreakSignal = {
+    name,
+    threshold: jailbreak_threshold,
+    include_history: formData.include_history || false,
+    description: formData.description || undefined,
+  }
+  if (method !== 'classifier') {
+    jailbreakEntry.method = method
+  }
+  if (formData.jailbreak_direction === 'response') {
+    jailbreakEntry.direction = 'response'
+  }
+  if (method === 'contrastive') {
+    const jailbreakPatterns = normalizeStringList(formData.jailbreak_patterns, 'Jailbreak patterns')
+    const benignPatterns = normalizeStringList(formData.benign_patterns, 'Benign patterns')
+    if (jailbreakPatterns.length > 0) jailbreakEntry.jailbreak_patterns = jailbreakPatterns
+    if (benignPatterns.length > 0) jailbreakEntry.benign_patterns = benignPatterns
+  }
+  return jailbreakEntry
+}
+
 export default function ConfigPageSignalsSection({
   config,
   isPythonCLI,
@@ -265,7 +385,7 @@ export default function ConfigPageSignalsSection({
     allSignals.push({
       name: jb.name,
       type: 'Jailbreak',
-      summary: `Method: ${method}, Threshold: ${jb.threshold}${jb.include_history ? ', includes history' : ''}`,
+      summary: `Method: ${method}, Threshold: ${jb.threshold}${jb.direction === 'response' ? ', scores the response' : ''}${jb.include_history ? ', includes history' : ''}`,
       rawData: jb,
     })
   })
@@ -698,57 +818,7 @@ export default function ConfigPageSignalsSection({
         ],
       })
     } else if (signal.type === 'Jailbreak') {
-      const fields: Array<{ label: string; value: React.ReactNode; fullWidth?: boolean }> = [
-        { label: 'Method', value: signal.rawData.method || 'classifier', fullWidth: true },
-        {
-          label: 'Threshold',
-          value: signal.rawData.threshold?.toString() || 'N/A',
-          fullWidth: true,
-        },
-        {
-          label: 'Include History',
-          value: signal.rawData.include_history ? 'Yes' : 'No',
-          fullWidth: true,
-        },
-      ]
-      if (signal.rawData.method === 'contrastive') {
-        fields.push(
-          {
-            label: 'Jailbreak Patterns',
-            value: (
-              <SignalStringListEditor
-                value={signal.rawData.jailbreak_patterns}
-                onChange={() => undefined}
-                addLabel=""
-                emptyLabel="No jailbreak patterns."
-                itemLabel="Jailbreak pattern"
-                readOnly
-              />
-            ),
-            fullWidth: true,
-          },
-          {
-            label: 'Benign Patterns',
-            value: (
-              <SignalStringListEditor
-                value={signal.rawData.benign_patterns}
-                onChange={() => undefined}
-                addLabel=""
-                emptyLabel="No benign patterns."
-                itemLabel="Benign pattern"
-                readOnly
-              />
-            ),
-            fullWidth: true,
-          },
-        )
-      }
-      fields.push({
-        label: 'Description',
-        value: signal.rawData.description || 'N/A',
-        fullWidth: true,
-      })
-      sections.push({ title: 'Jailbreak Signal', fields })
+      sections.push({ title: 'Jailbreak Signal', fields: jailbreakDetailFields(signal.rawData) })
     } else if (signal.type === 'PII') {
       sections.push({
         title: 'PII Signal',
@@ -857,11 +927,7 @@ export default function ConfigPageSignalsSection({
       easy_candidates: [],
       composer_operator: 'AND',
       composer_conditions: [],
-      jailbreak_threshold: 0.65,
-      jailbreak_method: 'classifier',
-      include_history: false,
-      jailbreak_patterns: [],
-      benign_patterns: [],
+      ...jailbreakFormDefaults(),
       pii_threshold: 0.5,
       pii_types_allowed: [],
       pii_include_history: false,
@@ -921,11 +987,7 @@ export default function ConfigPageSignalsSection({
             easy_candidates: [...(signal.rawData.easy?.candidates || [])],
             composer_operator: signal.rawData.composer?.operator || 'AND',
             composer_conditions: [...(signal.rawData.composer?.conditions || [])],
-            jailbreak_threshold: signal.rawData.threshold ?? 0.65,
-            jailbreak_method: signal.rawData.method || 'classifier',
-            include_history: !!signal.rawData.include_history,
-            jailbreak_patterns: [...(signal.rawData.jailbreak_patterns || [])],
-            benign_patterns: [...(signal.rawData.benign_patterns || [])],
+            ...jailbreakFormStateFrom(signal.rawData),
             pii_threshold: signal.rawData.threshold ?? 0.5,
             pii_types_allowed: [...(signal.rawData.pii_types_allowed || [])],
             pii_include_history: !!signal.rawData.include_history,
@@ -1213,30 +1275,10 @@ export default function ConfigPageSignalsSection({
           break
         }
         case 'Jailbreak': {
-          const jailbreak_threshold = formData.jailbreak_threshold ?? 0.65
-          if (jailbreak_threshold < 0 || jailbreak_threshold > 1) {
-            throw new Error('Jailbreak threshold must be between 0.0 and 1.0.')
-          }
-          const method = formData.jailbreak_method || 'classifier'
-          const jailbreakEntry: JailbreakSignal = {
-            name,
-            threshold: jailbreak_threshold,
-            include_history: formData.include_history || false,
-            description: formData.description || undefined,
-          }
-          if (method !== 'classifier') {
-            jailbreakEntry.method = method
-          }
-          if (method === 'contrastive') {
-            const jailbreakPatterns = normalizeStringList(
-              formData.jailbreak_patterns,
-              'Jailbreak patterns',
-            )
-            const benignPatterns = normalizeStringList(formData.benign_patterns, 'Benign patterns')
-            if (jailbreakPatterns.length > 0) jailbreakEntry.jailbreak_patterns = jailbreakPatterns
-            if (benignPatterns.length > 0) jailbreakEntry.benign_patterns = benignPatterns
-          }
-          newConfig.signals.jailbreak = [...(newConfig.signals.jailbreak || []), jailbreakEntry]
+          newConfig.signals.jailbreak = [
+            ...(newConfig.signals.jailbreak || []),
+            buildJailbreakEntry(name, formData),
+          ]
           break
         }
         case 'PII': {